Re: [PATCH] clone: teach --single-branch and --branch during --recurse

From: Emily Shaffer <emilyshaffer@google.com>
To: Jeff King <peff@peff.net>
Cc: git@vger.kernel.org
Subject: Re: [PATCH] clone: teach --single-branch and --branch during --recurse
Date: Thu, 16 Jan 2020 14:38:00 -0800	[thread overview]
Message-ID: <20200116223800.GM181522@google.com> (raw)
In-Reply-To: <20200109081150.GC3978837@coredump.intra.peff.net>

On Thu, Jan 09, 2020 at 03:11:50AM -0500, Jeff King wrote:
> On Wed, Jan 08, 2020 at 03:19:00PM -0800, Emily Shaffer wrote:
> 
> > Subject: clone: teach --single-branch and --branch during --recurse
> 
> A minor nit, but this subject confused me for a moment. I think we'd
> usually say "teach" for a new feature being implemented, and this is
> really just about passing along the existing features. Something like:
> 
>   clone: pass --single-branch and --branch when recursing submodules
> 
> would have been a bit more obvious (to me anyway).
> 
> > Previously, performing "git clone --recurse-submodules --single-branch"
> > resulted in submodules cloning all branches even though the superproject
> > cloned only one branch. Pipe --single-branch and its friend, --branch,
> > through the submodule helper framework to make it to 'clone' later on.
> 
> Since I don't really use submodules, I don't have much data or even
> intuition to go on. But could this be a regression for some situations?
> E.g., imagine I have a repo "parent" whose branch "foo" has a submodule
> "child", but "child" only has a branch "bar". What happens now if I "git
> clone --recurse-submodules --single-branch -b foo parent", and what will
> happen after this patch?
> 
> I think it works before, but doesn't now.
> 
> Setting up like this:
> 
>   git init child
>   (
>   	cd child
>   	git checkout -b bar
>   	echo whatever >file
>   	git add file
>   	git commit -m 'child commit'
>   )
>   
>   git init parent
>   cd parent
>   git checkout -b foo
>   git submodule add $PWD/../child
>   git commit -m 'add submodule'
> 
> if I use the current tip of master, I get:
> 
>   $ git clone --recurse-submodules --single-branch -b foo parent cur
>   Cloning into 'cur'...
>   done.
>   Submodule 'child' (/home/peff/tmp/parent/../child) registered for path 'child'
>   Cloning into '/home/peff/tmp/cur/child'...
>   done.
>   Submodule path 'child': checked out 'b5cbfcc9fec3b7d67e305468624fed2ba1aa4758'
> 
>   $ git -C cur/child log -1 --oneline | cat
>   b5cbfcc child commit
> 
> with your patch, I get:
> 
>   $ git.compile clone --recurse-submodules --single-branch -b foo parent new
>   Cloning into 'new'...
>   done.
>   Submodule 'child' (/home/peff/tmp/parent/../child) registered for path 'child'
>   Cloning into '/home/peff/tmp/new/child'...
>   warning: Could not find remote branch foo to clone.
>   fatal: Remote branch foo not found in upstream origin
>   fatal: clone of '/home/peff/tmp/parent/../child' into submodule path '/home/peff/tmp/new/child' failed
>   Failed to clone 'child'. Retry scheduled
>   Cloning into '/home/peff/tmp/new/child'...
>   warning: Could not find remote branch foo to clone.
>   fatal: Remote branch foo not found in upstream origin
>   fatal: clone of '/home/peff/tmp/parent/../child' into submodule path '/home/peff/tmp/new/child' failed
>   Failed to clone 'child' a second time, aborting
> 
>   $ git -C new/child log -1 --oneline | cat
>   11acb3a add submodule
> 
> (there's nothing checked out in the submodule).
> 
> I have no idea how common this kind of thing would be, and I expect in
> most cases your patch would do what people want. But we might need to be
> better about retrying without those options when the first clone fails.

Yeah, that's interesting. A retry sounds like a pretty solid approach,
although if someone's being cautious and using --single-branch I wonder
if that's really something they want (since that's avoiding grabbing
extraneous branches).

I suppose at the very least, --single-branch without --branch should
become recursive. Whether --branch should become recursive, I'm not
totally sure.

The two scenarios I see in conflict are this:
- Superproject branch "foo"
  - submodule A branch "foo"
  - submodule B branch "foo"
and
- Superproject branch "foo"
  - submodule A branch "bar"
  - submodule B branch "baz"

If we propagate --branch, the first scenario Just Works, and the second
scenario requires something like:

 git clone --recurse-submodules=no --single-branch --branch foo https://superproject
 git submodule update --init --single-branch --branch bar A/
 git submodule update --init --single-branch --branch baz B/

(I guess if the superproject knows what branch it wants for all the submodules,
you could also just do "git submodule update --init --single-branch" and
have it go and ask for all of them.)

If we don't propagate --branch, the second scenario Just Works, and the
first scenario requires something like:

 git clone --recurse-submodules=no --single-branch --branch foo https://superproject
 git submodule update --init --single-branch --branch foo

(That is, I think as long as 'update' takes --branch, even if it's not
passed along by 'clone', it should still work OK when delegating to
everyone.)

Let me know if I misunderstood what you were worried about.

I don't use submodules heavily for myself either. I'll try and ask
around a little to see what folks want, at least here. The ergonomics
seem pretty similar, so I guess it comes down to having explicit
documentation.

> 
> -Peff